Why Jev treats AI as a database, not a coworker — and

The launch. TypeSafe CEO Diogo Almeida releases Jev, a model built from a different premise: treat AI as infrastructure for code, not a chat interface. Jev outputs structured primitives (choice, score, no) rather than free text, aiming to replace string-based LLM outputs with predictable, machine-readable decisions. The target is any programmatic workflow where reliability and cost efficiency matter more than conversational fluency.

The core claim. Most AI today over-optimizes for human interaction, producing brittle, overconfident chat personalities unsuited for automated decisions. Jev optimizes for 'intelligence per dollar,' with built-in calibration scores that let developers set confidence thresholds. The model is designed as 'system one' — fast, cheap, and reliable on well-defined decision tasks — rather than the 'system two' reasoning chains that dominate current frontier models.

The underlying critique. Almeida argues that RLHF and safety refusal patterns actively harm programmatic use. A safety refusal that is a nuisance in chat becomes a production-breaking failure in an automated pipeline. He draws a sharp line: 'intelligence will be more like a database than a coworker.' AI safety alignment as practiced today fractures model capability, making it less reliable for the tasks software engineers actually need automated.

The organizational impact. If Jev delivers on its claims, teams evaluating AI for production workflows must reconsider their model selection criteria. Raw benchmark scores become irrelevant; workflow-specific evaluation, calibration, and cost-per-intelligence replace them. Engineering leaders will face pressure to move from general-purpose chat models to specialized, cheaper programmatic models for decision tasks, potentially splitting their AI stack.

The trade-offs and risks. Jev excels at simple, high-volume decisions but explicitly fails at multi-hop reasoning and complex agentic chains. Almeida acknowledges this limitation, framing it as a feature: use the right tool for the right task. The company operates without public benchmarks, relying on developer 'vibes' and internal evals. This trust-based approach carries adoption risk for compliance-heavy enterprises requiring third-party validation.

The practical signal for teams. The most immediate takeaway is architectural: decompose AI workflows into many small, verifiable decisions rather than one large prompt. Jev's structured output types (choice, score, no) map directly to programming primitives, enabling AI calls that are as predictable as function calls. Teams building automated pipelines should evaluate whether their current LLM usage can be replaced by cheaper, more reliable programmatic models for the 80% of simple decisions.